Setup

Python dependencies:

pip install -r requirements.txt

Usage

Examples can be run as follows:

deepspeed --num_gpus [number of GPUs] test-[model].py

NOTE: Local CUDA graphs for replaced SD modules will only be enabled when mp_size==1.

Example Output

Command:

deepspeed --num_gpus 1 test-stable-diffusion.py

Output:

./baseline.png
./deepspeed.png